Slack Code puts four rival coding agents in one channel — and doesn't sell any of them

What shipped

Tag a coding agent in a Slack conversation and it opens a code channel — a dedicated space for that one piece of work. Inside it, the agent posts code diffs that render inline and live HTML previews you can click. Enterprise review workflows route high-stakes changes to a human approver without anyone leaving the channel. Members can pause, redirect or stop the agent mid-task. When the work is done the channel auto-archives, staying searchable as a record of what was asked, what was built, and who approved it.

A new Agents & Tools tab collects agent conversations across threads, DMs and code channels into one dashboard, which is a small thing that matters — it is the difference between agents scattered through a workspace and agents you can actually account for.

The launch partners each use the code channel APIs differently. Claude can spin up a code channel from an existing thread and post a summary back to it. Devin responds in context inside team conversations and generates live previews. GitHub Copilot is pitched at the case where a non-technical colleague requests a fix and engineers review it in real time. Vercel posts a shareable live preview the moment a change deploys. ChatGPT is listed as coming soon — a real gap for teams standardised on Codex.

Conspicuously not on the list: Cursor, which is running its own strategy around always-on cloud agents billed against its subscription.

The format fits a specific queue

The honest case for Slack Code is narrower than “AI coding is now a team sport,” and it is better for being narrow.

Every engineering team has a shallow queue: the bug a support colleague noticed, the copy that needs changing on a pricing page, the small feature someone can describe precisely but not implement. That work is expensive not because it is hard but because of the handoffs — it gets described in Slack, translated into a ticket, picked up in an IDE, reviewed in a pull request, and reported back in Slack. Four context switches for twenty minutes of work.

Slack Code collapses that. The description, the implementation, the review and the record all happen where the request originated. Slack’s own number — over 70% of code channels open and close within a single day, from idea to merged PR — is a vendor statistic describing the tasks that suit the format, not a claim about all engineering work. Read that way, it is credible and it is the point.

What it is not is a replacement for a real harness. Work that requires holding a large codebase in context still belongs in an IDE or a terminal agent, where the working set is your repository rather than a channel’s scrollback. The realistic pattern is both, and teams evaluating that split will find the coding tools comparison and the Cursor vs Claude Code breakdown more useful than any single-vendor pitch.

The aggregator move

Here is the part worth sitting with. Slack built a container and shipped four competitors’ agents into it on equal terms, through published APIs, and charges nothing for the container.

That is a recognisable strategy, and it is the same one playing out one layer down this week. Stripe just paid a reported $7.5 billion for OpenRouter to own the neutral layer between applications and models, while Ramp launched a free competitor to the same layer on the same day. Slack is running the equivalent play between teams and agents: be the shell, be free, be where the work already is, and let the vendors compete inside your walls.

For buyers, aggregation cuts both ways.

The good side is genuine. If agents plug into a standard container through the same APIs, swapping one for another is a change of who you tag rather than a change of workflow — the same portability argument that makes model-agnostic orchestration valuable. Slack has no incentive to favour any one agent, because it sells none of them. That is a more credible neutrality than a lab or an IDE vendor can offer.

The other side is the bill. A container that hosts four agents does not consolidate four subscriptions into one; it makes running all four easy. The friction that used to stop a team from buying Claude Code and Devin and Copilot at once was partly that nobody wanted to juggle three tools. Remove the juggling and the natural equilibrium is more vendors, not fewer. If you are the person who signs off on AI spend, the free channel is the cheapest part of what you are about to authorise — a dynamic worth watching in any developer tooling budget.

The failure mode nobody markets

Slack has built the safety mechanics properly. Approvals route to named humans. Agents can be stopped mid-task. Everything archives into an audit trail. Compared with an agent operating in a terminal where only one person ever sees the output, a code channel is strictly more observable.

But observability is not scrutiny, and the failure mode here is cultural rather than technical: a thumbs-up in a busy channel is not a code review.

Pull requests work partly because of ceremony. There is a named reviewer, a request that sits in someone’s queue until they act, and a social cost to approving something you did not read. A channel inverts all three. Review becomes ambient, responsibility diffuses across everyone watching, and approval becomes a reaction emoji. Move code review from a pull request into a chat channel and you may trade rigour for visibility while feeling like you gained both.

The mitigation is unglamorous and effective: keep the merge gate exactly where it already is. Let the code channel be the place work gets described, built, previewed and discussed. Let the existing pull request review be the place it gets approved. Slack’s enterprise approval routing supports this — it is a control you configure, not a default you inherit. Teams that treat the channel as a faster path to a review, rather than a replacement for one, get the handoff savings without the governance loss.

What this means for the agent market

Slack Code is the clearest sign yet that coding agents are becoming interchangeable components inside someone else’s container rather than destinations users go to.

That is bad news for any agent whose moat was its interface and good news for the ones whose moat is capability. It also explains the absences. Cursor is not in the launch list because Cursor’s product is the environment — plugging into Slack’s container means competing on model quality alone, inside a room it does not control. OpenAI is coming soon rather than live, which for the largest player is a choice about sequencing as much as engineering.

The pattern holds across the week. Claude Tag put one Claude in a channel in June — a vendor shipping into Slack. Slack Code is Slack shipping a standard that vendors ship into. The second is a much stronger position, and it is the one that determines whose margin gets compressed over the next year.

The roadmap says the rest out loud: the code channel APIs will extend to marketing, legal review and IT. Which means the valuable thing being built here was never a coding feature. It is a standard container for agent work inside the application most companies already leave open all day, and the coding agents are simply the first tenants willing to prove it works.

Questions or corrections? Email Pick Right. Want the full list? See all news.